Oscillations in an artificial neural network convert competing inputs into a temporal code

Area of Science:

• Computational Neuroscience
• Machine Learning
• Computer Vision

Background:

• Computer vision (CV) often models the primate visual system, influencing artificial neural networks (ANNs) like convolutional neural networks (CNNs).
• However, ANNs typically ignore oscillatory dynamics observed in biological visual systems.
• Computational models of brain dynamics seldom incorporate CV principles.

Purpose of the Study:

• To integrate oscillatory dynamics from computational neuroscience into a simple ANN.
• To investigate if these dynamics can resolve input bottlenecks in ANNs.

Main Methods:

• A simple ANN was trained to classify individual letters.
• Temporal dynamics, including unit refraction and alpha-like oscillatory inhibition, were added to the hidden layer post-training.
• The network's performance was evaluated on single and dual letter classification tasks.

Main Results:

• The network without dynamics produced mixed outputs for simultaneous letters, indicating a bottleneck.
• Introducing oscillatory inhibition enabled sequential activation of output nodes for dual stimuli.
• The timing of sequential activation was controlled by the phase of the inhibitory oscillations.

Conclusions:

• Inhibitory oscillations can effectively segregate competing inputs in time within ANNs.
• This approach offers a novel method for improving ANN performance on complex tasks.
• The findings suggest potential applications in deeper network architectures and advanced machine learning problems.